I run a fleet of always-on agents across my Macs, and I drive the whole thing from my phone. One repo configures every machine. I clone it, paste a single prompt into an agent, and it works out which Mac it’s on and brings that machine up to spec from a runbook. The repo is their shared memory - each session reads it, does its work, and writes back what changed.
The roles split cleanly. My daily driver is hands-on: dev work, plus an agent that holds my signed-in browser, so I can hand it a task and watch it run from my phone. A home machine stays on around the clock and runs a few lanes in parallel
- each in its own account with its own desktop - so several agents work at once and I hop between them like tabs. A third box takes the unattended jobs: scheduled routines, and the occasional task that means driving a native app with no API, clicking through it the way a person would.
Building it changed which boundary I care about. For years the line was personal versus work. Now the lines that matter are attended versus unattended, and reversible versus irreversible. Once an agent can act on its own while I’m away, that’s the real distinction.
So safety is a gate, not a wall. I let agents hold real logins and real capability, then put a human confirmation step on the few actions that submit, pay, or delete - rather than sandboxing them into uselessness.
The other bet: build the substrate to outlast the model. Most of this is vendor-agnostic plumbing - accounts, sessions, a private network, plain-text runbooks a person and an agent can both read. Only a thin layer is tied to any one model. The orchestration is the durable asset, not the CLI.
What it changes for one person is the interesting part. A few always-on agents, run from a phone, is parallel workstreams without more headcount. That still feels like a cheat code.
What surprised me is how personal it all is. The runbooks are really just my accounts, my habits, the handful of actions I want to stop and confirm. Someone else’s would look nothing like mine.
It’s made me curious whether this is really an OS-level thing. I’m all-in on Apple, and they already have the parts - the accounts, the devices, the biometrics, the network between them. I keep wondering what a native version would feel like: an agent that holds your logins, works while you’re away, and checks in before anything you can’t undo. Though half the fun of mine is that it’s plain text I can poke at and rewire, and I’m not sure that survives becoming a product.